From Cysts to Metabolism: Trading Metaphors in Medical Language for the Polycystic Ovary Syndrome to Polyendocrine Metabolic Ovarian Syndrome Shifts.

Pneumonia is a major infectious disease worldwide and remains a leading cause of morbidity and mortality. Accordingly, standardized diagnosis, treatment, and antimicrobial management are essential components of modern health care. In China, the burden of community-acquired and hospital-acquired pneumonia remains substantial, driven in part by rapid population aging and increasing numbers of immunosuppressed patients and patients with multimorbidity []. At the same time, the evolution of bacterial resistance continues to outpace the development and clinical availability of new antimicrobials, making inappropriate antibiotic use a major global public health challenge []. In real-world practice, inappropriate antimicrobial prescribing remains common and may compromise treatment efficacy, increase the risk of adverse drug events, accelerate the spread of antimicrobial resistance, and raise health care costs []. For hospitalized patients with pneumonia, antibiotic selection often requires early empirical treatment before definitive microbiological results become available. At the same time, patient-specific factors such as hepatic or renal impairment, older age, and multimorbidity can substantially influence antibiotic selection and dose recommendation. These challenges highlight the need for clinical decision support (CDS) tools that can assist individualized antimicrobial therapy while remaining aligned with guideline constraints. Before the advent of large language models (LLMs), biomedical informatics had long explored CDS approaches. Early rule-based and knowledge-based expert systems [,] encoded medical knowledge as “if-then” rules with strong interpretability but relied heavily on manual maintenance and were difficult to update in response to evolving guidelines, emerging evidence, and complex clinical contexts. Subsequently, traditional natural language processing and machine learning methods were applied to tasks such as clinical text structuring, risk prediction, and adverse event detection [-]. Deep learning models, including recurrent neural networks and bidirectional encoder representations from transformers, further improved clinical text representations [-]. However, prediction or classification remained the dominant paradigm at this stage, often without providing an auditable chain of clinical reasoning [,,]. In the area of medication recommendation and individualized prescribing, previous studies have mainly focused on medication information extraction, prescription review alerts, or medication prediction [,,], with limited ability to jointly incorporate guideline constraints, patient-specific variation, and cross-modal clinical evidence. In addition, limited interpretability and traceability have remained major barriers to clinical adoption in high-risk medication decisions. In recent years, advances in LLMs for medical context understanding and text generation have expanded biomedical informatics from “structuring and prediction” toward “clinical text reasoning and explainable generation” [,]. These models have shown promise in tasks such as medical record information extraction and clinical question answering [-]. For medication recommendation, LLMs can integrate illness descriptions, prior medications, and test results to generate candidate regimens and rationales [-]. With retrieval-augmented generation, external sources such as guideline statements, drug labels, and local clinical pathways can be incorporated into the reasoning process, thereby improving knowledge coverage and output consistency [-]. However, in high-risk settings such as antibiotic selection and dose recommendation, important safety challenges remain. Adherence to fine-grained constraints, including dose limits, contraindications, and hepatic or renal dose adjustment, is often unstable []. Model conclusions may also vary according to input phrasing and context organization [], and hallucinations may produce potentially unsafe recommendations []. To support safe antimicrobial decision-making, LLM outputs should be grounded in traceable evidence, guided by explicit clinical constraints, and accompanied by reasoning that can be reviewed and verified by clinicians. These limitations highlight the need for constrained, traceable, and auditable LLM-based frameworks that can systematically incorporate external evidence and explicit clinical rules into the reasoning process.